AI Magazine Volume 25 Number 1 (2004) (© AAAI)
Articles
Annotating Protein
Function through
Lexical Analysis
Rajesh Nair and Burkhard Rost
■ We now know the full genomes of more than 60
organisms. The experimental characterization of
the newly sequenced proteins is deemed to lack behind this explosion of naked sequences (sequencefunction gap). The rate at which expert annotators
add the experimental information into more or
less controlled vocabularies of databases snails
along at an even slower pace. Most methods that
annotate protein function exploit sequence similarity by transferring experimental information for
homologues. A crucial development aiding such
transfer is large-scale, work- and management-intensive projects aimed at developing a comprehensive ontology for gene-protein function, such as
the Gene Ontology project. In parallel, fully automatic or semiautomatic methods have successfully
begun to mine the existing data through lexical
analysis. Some of these tools target parsing controlled vocabulary from databases; others venture
at mining free texts from MEDLINE abstracts or
full scientific papers. Automated text analysis has
become a rapidly expanding discipline in
bioinformatics. A few of these tools have already
been embedded in research projects.
P
roteins are the machinery of life. The information for life is stored in a four-letter
alphabet in the genes (deoxyribonucleic
acid [DNA]) (Alberts et al. 1994; Lodish et al.
2000). This four-letter DNA alphabet is translated into a 20-letter amino acid alphabet constituting the basic language for proteins, the
machinery of life. Proteins are formed by joining amino acids through peptide bonds. Proteins differ greatly in the number of amino
acids joined (from 30 to more than 30,000)
and the arrangement and types of amino acids
used (dubbed residues when joined in proteins). Proteins are the macromolecules that
perform all important tasks in organisms, such
as catalysis of biochemical reactions, transport
of nutrients, and recognition and transmission
of signals. The plethora of role aspects of any
particular protein is referred to as its function.
However, protein function is not a well-defined term. Instead, function is a complex phenomenon that is associated with many mutually overlapping levels: chemical, biochemical,
cellular, organism mediated, developmental,
and physiological. These levels are related in
complex ways; for example, protein kinases
can be related to different cellular functions
(such as cell cycle) and to a chemical function
(transferase) plus a complex control mechanism by interaction with other proteins. The
same kinase can also be the culprit that leads
to misfunction, or disease. Thus, identifying
protein function is a step toward understanding diseases and identifying drug targets (Brutlag 1998).
The first entire genome (DNA) sequence of a
free living organism, Haemophilus influenzae,
was published in 1995 (Fleischmann et al.
1995). Currently, we know the full genomes
for more than 100 organisms; for more than 60
of these, the data are publicly available and
contribute about 250,000 protein sequences,
that is, about one-fourth of all currently
known protein sequences (Carter et al. 2003;
Liu and Rost 2001; Pruess et al. 2003). The
number of entirely sequenced genomes is expected to continue growing exponentially for
at least the next few years. This explosion of se-
Copyright © 2004, American Association for Artificial Intelligence. All rights reserved. 0738-4602-2004 / $2.00
SPRING 2004
45
Articles
ID MYOD_HUMAN STANDARD;
PRT; 319 AA.
AC P15172;
...
DE Myoblast determination protein 1 (Myogenic factor MYF-3).
GN MYOD1 OR MYF3.
OS Homo sapiens (Human).
OC Eukaryota; Metazoa; Chordata; Craniata; Vertebrata; Euteleostomi;
OC Mammalia; Eutheria; Primates; Catarrhini; Hominidae; Homo.
OX NCBI_TaxID=9606;
...
CC -!- FUNCTION: INVOLVED IN MUSCLE DIFFERENTIATION (MYOGENIC FACTOR).
CC
INDUCES FIBROBLASTS TO DIFFERENTIATE INTO MYOBLASTS. ACTIVATES
CC
MUSCLE-SPECIFIC PROMOTERS. INTERACTS WITH AND IS INHIBITED BY THE
CC
TWIST PROTEIN. THIS INTERACTION PROBABLY INVOLVES THE BASIC
CC
DOMAINS OF BOTH PROTEINS (BY SIMILARITY).
CC -!- SUBUNIT: EFFICIENT DNA BINDING REQUIRES DIMERIZATION WITH ANOTHER
CC
BHLH PROTEIN. SEEMS TO FORM ACTIVE HETERODIMERS WITH ITF-2.
CC -!- SUBCELLULAR LOCATION: Nuclear.
CC -!- SIMILARITY: BELONGS TO THE BASIC HELIX-LOOP-HELIX (BHLH) FAMILY OF
CC
TRANSCRIPTION FACTORS. “MYOGENIC FACTORS” SUBFAMILY.
...
DR SWISS-2DPAGE; GET REGION ON 2D PAGE.
KW Myogenesis; Differentiation; Developmental protein; Nuclear protein;
KW Transcription regulation; DNA-binding.
FT DNA_BIND 109 121
BASIC DOMAIN.
FT DOMAIN
122 161
HELIX-LOOP-HELIX MOTIF (BY SIMILARITY).
FT CONFLICT 124 124
K -> E (IN REF. 2).
...
Figure 1. Protein Entry in SWISS-PROT.
The SWISS-PROT identifier for the protein MYOD_HUMAN is found under the header ID. The type of protein and its source organism are
found under the DE and OS headers, respectively. Detailed functional information regarding the protein is found under the header CC.
This information is written in plain English and is not suitable for computer analysis. Following the KW header are keywords describing
the function of the protein. The keywords use a restricted vocabulary and are ideal for tools for text analysis.
quence information has widened the gap between the number of protein sequences deposited in public databases and the experimental characterization of the corresponding
proteins (Baker and Brass 1998; Koonin 2000;
Lewis et al. 2000; Rost and Sander 1996). Bioinformatics, sometimes referred to as functional
genomics, plays a central role in bridging the sequence-function gap through the development
of tools for faster and more effective prediction
of protein function (Bork et al. 1998; Fleischmann et al. 1999; Luscombe et al. 2001; Valencia 2002; Valencia and Pazos 2002).
Here, we briefly review some of the attempts
at annotating function through homology
transfer. The most widely used methods that
allow guessing protein function rely on the
ability to correctly mine the information deposited in public databases and scientific journals. Although we need a comprehensive ontology for protein function, developers have
46
AI MAGAZINE
begun to successfully explore the marvels of an
ever-increasing body of research in biology and
medicine. There are two major methods attempting automatic lexical analysis: (1) parsing of controlled vocabulary from databases
and (2) mining unstructured text available
from scientific publications. We could not cover all the promising approaches that have
mushroomed over the last 5 to 10 years; therefore, we focus in detail on a few success stories.
Annotations and Annotation
Transfer of Protein Function
Information about protein sequences is stored
in public databases such as SWISS-PROT and
TREMBL (table 1). SWISS-PROT (Apweiler 2001;
Bairoch and Apweiler 2000) is an expert-curated database that also contains annotations
about function (figure 1). These annotations
Articles
Database
SWISS-PROT
URL
(Boeckmann et al. 2003)
www.ebi.ac.uk/swissprot
TrEMBL (Boeckmann et al. 2003)
www.ebi.ac.uk/trembl
Gene Ontology (Ashburner et al. 2000)
www.geneontology.org
MIPS
PEP
(Mewes et al. 2000)
(Carter et al. 2003)
mips.gsf.de
cubic.bioc.columbia.edu/db/PE1
Table 1. Web Sites of Major Databases and Genome Resources.
are added by a team of expert annotators who
extract this information primarily from journal
publications (Junker et al. 2000). TREMBL
(Bairoch and Apweiler 2000) consists of entries
that are derived from the translation of all coding sequences in the EMBL nucleotide sequence
database that are not in SWISS-PROT. Unlike
SWISS-PROT records, those in TREMBL are awaiting manual annotation. SWISS-PROT currently
contains only 113,470 sequence entries, and
the TREMBL database contains over 755,169 sequence entries (Boeckmann et al. 2003).
Annotations of function primarily occur
through homology transfer. Experimentally
determining protein function continues to be
a laborious task that can take enormous resources; for example, more than a decade after
the discovery, we still do not know the precise
and entire functional role of the prion protein
(Harrison et al. 1997). The automatic elucidation of the protein function is therefore an appealing challenge (Apweiler et al. 1997; Eisenberg et al. 2000; Gaasterland and Sensen 1996).
The most commonly used approach for the automatic elucidation of protein function relies
on the fact that two proteins with similar sequence often have a similar function.
The basic idea to exploit this fact involves
the following steps: (1) extract the experimental information from the literature into a controlled vocabulary of annotated databases; (2)
establish thresholds T for pairwise sequence
similarity that imply similarity in function; (3)
for a protein U of unknown function, search
the database for proteins {K} that have a sequence similarity to U: SIM(K, U) > T; and (4) if
any such protein K is found, transfer its annotation to U. Albeit this concept appears
straightforward, in practice, there are many
hurdles to overcome: First, the function is not
well defined; hence, it is very difficult to create
controlled vocabularies (Ashburner et al. 2000;
Bairoch and Apweiler 2000). Second, because
function is such a complex phenomenon, it is
very difficult to assign one number that describes all these roles at once (Ashburner et al.
2000; Todd et al. 2001). Third, to add to the
complication, it seems that the precise values
for thresholds of significant sequence similarity (T) are actually specific to particular function—that is, become T(F)—and have to be
reestablished for any given task (Devos and Valencia 2000; Nair and Rost 2002a, 2002b;
Ouzounis et al. 1998; Rost 2002, 1999; Todd et
al. 2001; Wilson et al. 2000; Wrzeszczynski and
Rost 2003). In general, the inference of function is reliable only for very high levels of sequence similarity (Devos and Valencia 2001;
Nair and Rost 2002; Rost 2002). For reliably inferring the subcellular localization of a protein
using homology transfer, a sequence identity
of more than 80 percent is required. At this sequence identity, subcellular localization annotations can be transferred at more than 90-percent accuracy. Below this threshold, the
accuracy of annotation transfer rapidly decreases (Nair and Rost 2002b).
Several pitfalls in transferring annotations of
function have been reported, for example, having inadequate knowledge of thresholds for
“significant sequence similarity”; using only
the best database hit; or ignoring the domain
organization of proteins (Bork and Koonin
1998; Devos and Valencia 2001; Doerks et al.
1998; Galperin and Koonin 2000). Despite all
these problems, the majority of annotations
about function in public databases result from
homology transfer (Devos and Valencia 2001;
Koonin 2000; Valencia 2002). Databases such
as SWISS-PROT usually do not provide pointers
for the origin of the information. One problem
arising from this approach is that it is difficult
to distinguish annotations that are experimental from those that are predicted.
Problem 1:
Multiple Levels of Description
The function of a protein is context depen-
SPRING 2004
47
Articles
dent. Database annotations of protein function
are often confusing because of the variety of
functional roles (Attwood 2000). For reliable
automatic predictions, computer-readable hierarchical descriptions of function are needed
(Bork et al. 1998; Overbeek et al. 1997).
Several groups and associations have ventured to introduce numeric schemata to define
function. The first attempt was the introduction of enzyme classification numbers (Webb
1992); this classification uses four digits to classify enzymatic activity. The first enzyme classification digit distinguishes the general types of
enzymes; the second enzyme classification digit specifies the substrate (oxireductases), the
group transferred (transferases), the type of
bond (hydrolases, lyases, ligases), or the type of
reorganization (isomerases). The third and
fourth digits provide more detail (for an excellent survey of structural aspects of enzymatic
function, see Todd, Orengo and Thornton
[2001]). MIPS attempts to extend this idea to a
wider perspective, with more proteins and
more roles, through its classification catalog
(Mewes et al. 2000).
Arguably, the most impressive gargantuan
effort at defining ontology for protein function
originates from the gene ontology consortium
(Ashburner et al. 2000). Gene ontology distinguishes three levels of protein function: (1)
molecular, (2) biological, and (3) cellular. At
the molecular level, the protein can, for example, catalyze a metabolic reaction and recognize or transmit a signal. In a biological
process, a set of many cooperating proteins is
responsible for achieving broad biological
goals, for example, mitosis, purine metabolism, or signal transduction cascades. The cellular category includes the structure of subcellular compartments, the localization of
proteins, and macromolecular complexes. Examples include nucleus, telomere, and origin
recognition complex. The subcellular localization of a protein is an essential attribute for
this level. The totality of the physiological subsystems and their interplay with various environmental stimuli determine properties of the
phenotype, the morphology and physiology of
the organism and its behavior. Gene ontology
is not complete. In fact, now after almost a
decade of efficient work, the first notable coverage of the experimental information is complete, and the developers contemplate restarting. Nevertheless, gene ontology constitutes
the best set of definitions available today.
Problem 2: No Machine-Readable
Functional Information
Nearly all databases present the protein se-
48
AI MAGAZINE
quence in formats that are more or less
straightforward for parsing by computers.
However, annotations are mostly written in
plain text using a rich biological vocabulary
that often varies in different areas of research
(figure 1). Such annotations are primarily
meant for the eyes of human experts; hence,
they are not machine readable (Eisenhaber and
Bork 1999). Another problem that hampers automatic annotations is the quality of database
annotations: Only a few database groups attempt a quality control of curated annotations
(Tsoka and Ouzounis 2000).
Automatic Lexical Analysis of
Controlled Vocabularies
Protein databases such as SWISS-PROT usually
contain functional annotations at a very detailed level of biochemical function; for example, a given sequence is annotated as a cdc2 kinase but not as being involved in intracellular
communication (Apweiler 2001; Tamames et
al. 1998). A number of text-analysis tools have
been implemented that infer various aspects of
cellular function from database annotations of
molecular function. Many methods explore
the functional annotations in SWISS-PROT, especially the keyword annotations (Eisenhaber
and Bork 1999; Fleischmann et al. 1999; Nair
and Rost 2002a; Tamames et al. 1998).
SWISS-PROT currently contains over 800 keyword functional descriptors. Semantic analysis
of the keywords is used to categorize proteins
into classes of cellular function (Andrade et al.
1999b; Bork et al. 1992; Karp et al. 1999;
Ouzounis et al. 1996; Riley 1993; Riley and
Labedan 1997). There are two types of methods: (1) fully automated and (2) semiautomatic.
With fully automated methods, the problem
of automatically extracting rules from keywords has parallels to the problem of text categorization, that is, assigning predefined categories to free-text documents. Many statistical
learning methods have been applied to this
problem, including nearest-neighbor classifiers
(Yang and Pederson 1997), multivariate regression models (Schutze et al. 1995; Yang and
Chute 1992), probabilistic Bayesian models
(Lewis and Ringuette 1994), symbolic rule
learning (Apte et al. 1994), and m-ary (multiple-category) classifiers such as the k–nearest
neighbor (Dasarathy 1991) and the linear least
squares fit (LLSF). These methods have been intensively studied and are among the most accurate for text categorization (Yang and Liu
1999). The majority of the tools for annotating
function are based on one of these methods.
Articles
Database
URL
LOCkey (Nair and Rost 2002)
cubic.bioc.columbia.edu/services/LOCke
GeneQuiz (Tamames et al. 1998)
jura.ebi.ac.uk:8765/ext-genequiz/
Meta_A (Eisenhaber and Bork 1999)
mendel.imp.univie.ac.at/CELL_LC
AbXtract (Andrade and Valencia 1998)
columba.ebi.ac.uk:8765/andrade/a
SUISEKI
(Blaschke and Valencia 2002)
www.pdg.enb.uam.es/suiseki/
Table 2. Resources for Text Analysis.
Some of the major methods for annotating
function are LOCKEY (Nair and Rost 2002),
SPEARMINT (Bazzan et al. 2002; Kretschmann et
al. 2001), and the support vector machine
(SVM)-based approach of Stapley et al. (2002)
(table 2).
Semiautomated methods are based on building dictionaries of rules. Keywords characteristic of each of the functional classes are first extracted from a set of classified example
proteins. With these keywords, a library of
rules is created that associates a certain pattern
of occurrence of keywords to a functional class.
The major methods in this category are EUCLID
(Tamames et al. 1998), META_A (Eisenhaber and
Bork 1999), and RULEBASE (Fleischmann et al.
1999) (table 2). We review the LOCKEY and EUCLID algorithms as examples of the two main
approaches.
The LOCKEY system (Nair and Rost 2002a) is
a novel m-ary classifier that predicts the subcellular localization of a protein based on SWISSPROT keywords. The algorithm can be divided
into two steps (figure 2): (1) building data sets
of trusted vectors for known proteins and (2)
classifying unknown proteins. First, a list of
keywords is extracted from SWISS-PROT for all
proteins with known subcellular localization.
On average, most proteins have between two
and five keywords. A data set of binary vectors
(Salton 1989) is generated for each protein by
representing the presence of a certain keyword
in the protein by 1 and absence by 0. Second,
to infer subcellular localization of an unknown
protein U, all keywords for U are read from
SWISS-PROT. These keywords are translated into a
binary keyword vector. From this original keyword vector, LOCKEY generates a set of all possible combinations of alternative vectors by
flipping vector components of value 1 (presence of keyword) to 0 in all possible combinations. For example, for a protein with three
keywords, there are 23 – 1 = 7 possible subvec-
tors: 111, 110, 101, 011, 100, 010, and 001.
These subvectors constitute all possible keyword combinations for protein U. The keyword
combination, that is, subvector, that yields the
best classification of U into 1 of 10 classes of
sub-cellular localizations is found. All exact
matches of each of the subvectors to any of the
proteins in the trusted set are retrieved by finding all proteins in the trusted set that contain
all the keywords present in the subvector. By
construction, the proteins retrieved in this way
can also contain keywords not found in U.
The next task is to estimate the surprise value of the given assignment. Toward this end,
LOCKEY simply compiles the number of proteins belonging to each type of subcellular localization. This procedure is repeated in turn
for each of the subvectors, and localization is
finally assigned to a protein by minimizing an
entropy-based objective function. The system
accurately solves the classification problem
when the number of data points (proteins) and
the dimensionality of the feature space (number of keywords) are not too large. LOCKEY
reached a level of more than 82-percent accuracy in a full cross-validation test. However, because of a lack of functional annotations, the
system failed to infer localization for more
than half of all proteins in the test set (Note: A
number of SWISS-PROT keywords are biologically
correlated to subcellular localization; for example, DNA-binding proteins always localize to
the nucleus. These keywords, which were all
simple one-to-one correlations, were excluded
from testing because the goal was to estimate
the true ability of the algorithm to infer complex correlations among the keywords). For
five entirely sequenced proteomes—(1) Saccharomyces cerevisiae (yeast), (2) Caenorhabditis elegans (worm), (3) Drosophila melanogaster (fly),
(4) Arabidopsis thaliana (plant), and (5) a subset
of all human proteins—the LOCKEY system automatically found about 8000 new annota-
SPRING 2004
49
Articles
SWISS-PROT
Unknown Query
Localization
Annotated
Find Homologues
Extract
Keywords
SUB vectors
0100110010..
0100100000..
0100100010..
.............
Find
Matching
Vector
Best Classifier
(by entropy)
Find Homologues
TRUSTED vectors
cyt:
00010110011..
01001010010..
....
nuc:
01100101001..
00110011001..
Infer Localization
for Query
Figure 2. The LOCKEY Algorithm.
A sequence-unique data set of localization-annotated SWISS-PROT proteins was first compiled. Keywords were
extracted for these proteins and merged with any keywords found in homologues. The keywords were represented as binary vectors in the trusted vector set. An unknown query was first annotated with keywords
through identification of SWISS-PROT homologues. Keywords for the query were represented as binary vectors.
All possible keyword combinations were constructed (the subvectors). The best-matching vector was found
based on entropy criteria (Nair and Rost 2002a). This vector was used to infer localization for the query.
tions about subcellular localization. LOCKEY
has been optimized to provide fast annotations. Annotating the entire C. elegans proteome took less than 4 hours on a Pentium III
900-megahertz machine. The algorithm is limited to problems with relatively few data points
(proteins) in the vector set (n << 1000000) and
with few keywords (n << 10000).
The EUCLID system (Tamames et al. 1998) uses SWISS-PROT keywords to classify proteins into
14 classes of cellular function according to the
scheme originally proposed by Monika Riley
(Karp et al. 1999; Krawiec and Riley 1990; Riley
1993; Riley and Labedan 1997). The 14 classes
are grouped into 3 broad functional classes: (1)
energy, (2) information, and (3) communication. The EUCLID system can be summarized as
followed: First, keywords characteristic of each
of the functional classes are extracted from a
set of classified example proteins provided by a
50
AI MAGAZINE
human expert. This dictionary of characteristic
keywords satisfies the following criteria: (1) only keywords with functional meaning are used,
and keywords with no functional information
are excluded (for example, hypothetical or
three-dimensional structure); (2) only keywords appearing in more than one SWISS-PROT
entry are considered; and (3) only keywords
with more than 85 percent of their occurrences
in a single functional category are included in
the dictionaries.
Next, to assign sequences to classes, a simple
voting scheme is used. A sequence is automatically classified in the functional class to which
the majority of its keywords belong. The dictionary of keywords is then used to automatically
assign all proteins from the database if a sufficient match is found. Proteins thus assigned to
a functional class are analyzed to extract a new,
more extensive dictionary of characteristic key-
Articles
ID: HXKB_YEAST
DE: HEXOKINASE
...
KW: KINASE, ATP-BINDING…
SWISS-PROT File
START
Since HXKB_YEAST is in the ENERGY class, increase counts of
keywords “kinase” and ATP-binding in the “Energy” class
STDL_YEAST
HXKB_YEAST
FTSI_ECOLI
RNP1_YEAST
BUD5_YEAST
KAPA_YEAST
→
→
→
→
→
→
…
…
ENERGY
ENERGY
INFO
INFO
COMM
COMM
KEYWORDS
ACTIVATOR
ATP-binding
KINASE
PHEROMONE
→
→
→
→
→
→
→
→
…
ENERGY
ENERGY
INFO
INFO
COMM
COMM
INFO
ENERGY
2
15
14
0
17
25
0
0
0
17
1
3
Count the Occurrences of Each
Keyword in Each Field
Set of Proteins Classified
by Human Experts
STDL_YEAST
HXKB_YEAST
FTSI_ECOLI
RNP1_YEAST
BUD5_YEAST
KAPA_YEAST
XXXX
YYYY
ENERGY INFO COMM
Iterate
KEYWORDS
ACTIVATOR
ATP-binding
KINASE
PHEROMONE
UBIQUINONE
CLASS
INFO
NONE
ENERGY
COMM
NONE
Generated Dictionary of Keywords
Assign each keyword to one class
Add the Classified Proteins
to the Initial Set
PROTEIN: XXXX
KEYWD: ACTIVATOR
ASSIGNED TO: INFO
PROTEIN: YYYY
KEYWD: KINASE, ATP-binding
ASSIGNED TO ENERGY
PROTEIN: ZZZZ
KEYWD: ACTIVATOR, KINASE
ASSIGNED TO: NONE
Assign Proteins to Classes
Figure 3. The EUCLID Algorithm.
Scheme of the iterative method used to classify sequences in three functional classes. The classification relies on the definition of a dictionary of keywords characteristic for a particular functional class. First, experts assign hexokinase from yeast (hxkb_yeast) to the energy
class. Second, a keyword dictionary is constructed scoring the keywords associated with hexokinase in the energy class. Third, the same
dictionary is then extended to classifying other proteins. The process is iterated until no more keywords are gained.
words. The process is iterated until the classification quality no longer increases (figure 3). A
limitation of this approach is that only simple
correlations between keywords can be discovered. The method is easily scalable and can be
applied to very large protein databases. For the
genome sequence of Mycoplasma genitalium
(Fraser et al. 1995), the EUCLID system was able to
classify 52 percent of the sequences at a classifi-
cation accuracy of 82 percent. The EUCLID algorithm has been incorporated into the GENEQUIZ
workbench for sequence analysis (Andrade et al.
1999a). GENEQUIZ is a semiautomated protein
sequence analysis workbench whose principal
purpose is to infer a specific and reliable functional assignment together with a broad cellular
role for a query protein by analyzing annotations from sequence database matches.
SPRING 2004
51
Articles
Mining Free Text
from the Literature
Experimental results are usually published first
in scientific journals. Because such publications do not conform to any standardized
rules, this information is not computer readable. At best, this lack of automation leads to a
severe delay in incorporating the information
into databases. Furthermore, a lot of the data
will lie buried in the literature forever. One solution to this problem will be to adopt standards similar to the PDB model for protein
structure for depositing functional information into databases, that is, requiring deposition of, say, protein-protein interactions into a
public database along with publication.
However, currently, mining text is the only
way of retrieving functional information from
the literature. In recent years, many groups
have worked on dedicated problems in this
area, such as machine selection of articles of interest (Iliopoulos et al. 2001; Shatkay et al.
2000), automated extraction of information
using statistical methods (Stapley and Benoit
2000; Stephens et al. 2001), and natural language–processing techniques (Friedman et al.
2001; Ng and Wong 1999; Thomas et al. 2000;
Yakushiji et al. 2001) as well as setting up specialized knowledge bases for storing molecular
knowledge (Stevens et al. 2000). The invaluable electronic availability of scientific publications through MEDLINE (Airozo et al. 1999)
has not only severely affected the ways of writing papers and doing science in general, it has
also enabled the development of an avalanche
of methods that mine these data. Automatic
text-analysis tools can assist human annotators
and can thus significantly shorten the time lag
of functional annotations. One of the most
crucial bottlenecks for automated text analysis
is the mapping of gene-protein names (Hatzivassiloglou et al. 2001; Valencia 2002). Although this problem might be overcome in the
near future by particular standards adopted by
journals, this hurdle currently hinders the
availability and usefulness of public methods
considerably.
Many tools focus on mining MEDLINE abstracts. Although the principal reason for this
restriction is supposedly related to complexity
(abstracts available fit onto a disk and can be
searched quickly), abstracts are occasionally
more easy to mine because many papers contain less precise and less well-supported sections in the text that are difficult for machines
to distinguish from more informative sections
(Andrade and Bork 2000; Ding et al. 2002;
Hersh et al. 1992).
52
AI MAGAZINE
The current version of MEDLINE contains
nearly 12 million abstracts and occupies approximately 43 gigabytes of disk space. A
prominent example of methods that target entire papers is still restricted to a small number
of journals (Friedman et al. 2001; Krauthammer et al. 2002). The task of unraveling information about function from MEDLINE abstracts can be approached from two different
angles. In the first approach, computational
techniques for understanding text written in
natural language are based on lexical, syntactical, and semantic analysis (Cowie and Lehnert
1996; Salton 1989). In addition to indexing
terms in documents, natural language–processing methods extract and index higher-level semantic structures composed of terms and relationships between terms, which can be done in
different ways (for general discussion, see
Baeza-Yates and Ribeiro-Neto [1999]). However, this approach is confronted with the variability, fuzziness, and complexity of human
language (Andrade and Bork 2000). The GENIES
system (Friedman et al. 2001; Krauthammer et
al. 2002) for automatically gathering and processing knowledge about molecular pathways
and the IFBP transcription-factor database (Ohta, Yamamoto et al. 1997) are natural language–processing-based systems. The second
approach and one that might be more relevant
in practice is based on the treatment of text
with statistical methods (Andrade and Valencia
1998; Yang 1996). In this approach, the possible relevance of words in a text is deduced from
a comparison of the frequency of different
words in this text with the frequency of the
same words in reference sets of text (Berry et al.
1995). Some of the major methods using the
statistical approach are ABXTRACT (Andrade and
Valencia 1998, 1997) and the automatic pathway discovery tool of Ng and Wong (1999)
(table 2).
There are advantages to each of these approaches (grammar or pattern matching). Generally, the less syntax that is used, the more domain specific the system is. Thus, a robust
system can be constructed relatively quickly,
but many subtleties can be lost in the interpretation of the sentence. In some applications,
however, the domain-dependent patternmatching approach might be the only way to
attain reasonable performance in the near future (Allen 1995).
The ABXTRACT system (Andrade and Valencia
1998, 1997; Blaschke 2001; Blaschke and Valencia 2001) is triggered by collections of abstracts
related to a given protein, and it is able to extract functional information directly from
MEDLINE abstracts. Relevant keywords are se-
Articles
lected by their relative accumulation in comparison with a domain-specific background distribution. To obtain a representative set of words
(and their abundance) in protein families, the
background distribution of abstracts is chosen
to represent the widest range of protein families.
For each of the representative set (dictionary) of
words, two statistical parameters are computed:
(1) their frequency in each family and (2) the
deviation of the distribution of their frequencies
in the set of families. Provided with a query
family and an associated set of MEDLINE abstracts, words that are likely to be functionally
important for the family (putative keywords)
are found by comparison with the background
set. This comparison is done by measuring the
frequency of the relevant word in the query
family relative to its background frequency of
occurrence using a z-score (Andrade and Valencia 1998). Words with a high z-score are likely to
be potential keywords for the family.
The system has been tested on a number of
different protein families and showed a good
ability to extract functionally important keywords. A modification of this algorithm, called
SUISEKI (system for information extraction on
interactions) (Blaschke and Valencia 2001;
Blaschke et al. 2002), has been applied to the
problem of extracting protein-protein interaction from MEDLINE abstracts. In addition to
the statistical approach of ABXTRACT, SUISEKI also takes advantage of the analysis of the syntactic structure of phrases and other developments in computational linguistics. The SUISEKI
system was able to extract almost 70 percent of
the interactions present in a relatively large
text corpus at approximately 80-percent accuracy for the best-defined interactions. The SUISEKI system discovered a total of 4657 proteinprotein interactions in cell-cycle proteins in
yeast from a corpus of approximately 5300 abstracts (approximately 12 megabytes).
The authors identify a number of sources of
error in mining MEDLINE abstracts: First is erroneous detection of protein names. There is
no systematic nomenclature for gene and protein names, which has led to a number of possible writing variants and synonyms being associated with the proteins that makes
detection and classification difficult. Second
are indirect references and anaphoric expression. This problem is key for the analysis of
MEDLINE abstracts, where protein names can
be given in the title or initial sentences and later treated with forms such as the protein or
mentioned as a general class of proteins such as
the kinase. Third are deficiencies in the information-extraction technology. Incorrect parsing of sentences as a result of limitations im-
posed by the parsers and the use of complicated sentence structures to convey results are
some other areas where the information-extraction applications require improvement.
Conclusions
What is to be expected from computational genomics in the near future? As we illustrated in
this article, our battery of tools is becoming increasingly sophisticated, and our ability to annotate protein function using computers is
generally improving. However, to fully exploit
genome information, we still need to progress
from methods derived mostly from traditional
sequence analysis that examine genome sequences individually to algorithms and databases that exploit the inherent properties of
entire genomes. The development of a standardized ontology is an important step in this
direction. Text-based tools such as LOCKEY, EUCLID, and the MET_A annotator that infer cellular function from detailed annotations of molecular function found in databases can be
useful aids in the development of ontologies.
The development of tools for the extraction
of useful functional information from the existing literature is still in its infancy. The complexity and fuzziness of natural language make
it extremely difficult for computer algorithms
to parse and extract useful information from
text. The development of a standardized vocabulary for reporting experimental discovery
in the scientific literature will go a long way toward simplifying the process of extracting information directly from literature.
Genomics-based drug discovery is heavily
dependent on accurate functional annotations. Toward this end, bioinformatics will
need to deliver highly integrated, interoperable
data “warehouses” that allow the user to reason over disparate data sources and ultimately
enable knowledge-based inference and innovation. The road toward satisfactory solutions
might be long. However, the first successes
have been encouraging. One important lesson
from the successes of bioinformatics over the
last decade continues to be that integrated
tools become successful when their developers
are integrated with the wet-lab biologists.
Acknowledgments
Thanks to the following members of our group:
Jinfeng Liu and Megan Restuccia, for computer
assistance, and Dariusz Przybylski and Kazimierz Wrzeszczynski, for helpful discussions.
The authors were supported by grant DBI0131168 from the National Science Foundation. Last, but not least, thanks to all those
SPRING 2004
53
Articles
who deposit their experimental data in public
databases and to those who maintain these
databases.
References
Airozo, D.; Allard, R.; Brylawski, B.; Canese, K.; Kenton, D.; Knecht, L.; Krasnov, S.; Sandomirskiy, V.;
Sirotinin, V.; Starchenko, G.; Wilbur, J.; and Zipser, J.
1999. MEDLINE. Bethesda, Md.: National Library of
Medicine.
Alberts, B.; Bray, D.; Roberts, K.; and Watson, J. 1994.
Molecular Biology of the Cell. New York: Garland.
Allen, J. 1995. Natural Language Understanding. New
York: Addison-Wesley.
Andrade, M. A., and Bork, P. 2000. Automated Extraction of Information in Molecular Biology. Federation of European Biochemical Societies Letters 476(1–2):
12–17.
Andrade, M. A., and Valencia, A. 1998. Automatic
Extraction of Keywords from Scientific Text: Application to the Knowledge Domain of Protein Families.
Bioinformatics 14(7): 600–607.
Andrade, M. A., and Valencia, A. 1997. Automatic
Annotation for Biological Sequences by Extraction of
Keywords from MEDLINE Abstracts. Development of
a Prototype System. In Proceedings of the Fifth International Conference on Intelligent Systems for Molecular
Biology, 25–32. Menlo Park, Calif.: AAAI Press.
Andrade, M.; Brown, A. N.; Leroy, C.; Hoersch, S.; de
Daruvar, A.; Reich, C.; Franchini, A.; Tamames, J.; Valencia, A.; Ouzounis, C.; and Sander, C. 1999. Automated Genome Sequence Analysis and Annotation.
Bioinformatics 15(5): 391–412.
Andrade, M. A.; Ouzounis, C.; Sander, C.; Tamames,
J.; and Valencia, A. 1999. Functional Classes in the
Three Domains of Life. Journal of Molecular Evolution
49(5): 551–557.
Apte, C.; Damerau, F.; and Weiss, S. 1994. Toward
Language-Independent Automated Learning of Text
Categorization Models. In Proceedings of the Seventeenth Annual ACM/SIGIR Conference, 24–30. New
York: Association of Computing Machinery.
Bairoch, A., and Apweiler, R. 2000. The SWISS-PROT
Protein Sequence Database and Its Supplement
TREMBL in 2000.” Nucleic Acids Research 28(1):
45–48.
Baker, P. G., and Brass, A. 1998. Recent Developments in Biological Sequence Databases. Current
Opinion in Biotechnology 9(1): 54–58.
Bazzan, A. L.; Engel, P. M.; Schroeder, L. F.; and Da
Silva, S. C. 2002. Automated Annotation of Keywords for Proteins Related to Mycoplasmataceae Using Machine Learning Techniques. Bioinformatics
18(Suppl. 2): S35–S43.
Berry, M. W.; Dumais, S. T.; and O’Brien, G. W. 1995.
Using Linear Algebra for Intelligent Information Retrieval. Society for Industrial and Applied Mathematics
Review 37(4): 573–595.
Blaschke, C. 2001. Applications of Information-Extraction Techniques to Molecular Biology. Ph.D.
diss., CNB, Madrid, University of Autonoma Madrid.
Blaschke, C., and Valencia, A. 2002. The Frame-Based
Module of the SUISEKI Information-Extraction System.
IEEE Intelligent Systems 17(1): 14–20.
Blaschke, C., and Valencia, A. 2001. The Potential
Use of SUISEKI as a Protein Interaction Discovery Tool.
Genome Informatics Series Workshop. Genome Informatics 12:123–134.
Blaschke, C.; Hirschman, L.; and Valencia, A. 2002.
Information Extraction in Molecular Biology. Briefings in Bioinformatics 3(2): 154–165.
Boeckmann, B.; Bairoch, A.; Apweiler, R.; Blatter, M.
C.; Estreicher, A.; Gasteriger, E.; Martin, M. J.; Michoud, K.; O’Donovan, C.; Phan, I.; Pibout, S.; and
Schneider, M. 2003. The SWISS-PROT Protein Knowledge Base and Its Supplement TREMBL in 2003. Nucleic Acids Research 31(1): 365–370.
Bork, P.; Dandekar, T.; Diaz-Lazcoz, Y.; Eisenhaber, F.;
Huynen, M.; and Yuan, Y. 1998. Predicting Function:
From Genes to Genomes and Back. Journal of Molecular Biology 283(4): 707–725.
Bork, P,. and Koonin, E. V. 1998. Predicting Functions from Protein Sequences—Where Are the Bottlenecks? Nature Genetics 18(4): 313–318.
Apweiler, R.; Gateau, A.; Contrino, S.; Martin, M. J.;
Junker, V.; O’Donovan, C.; Lang, F.; Mitaritonna, N.;
Kappus, S.; and Bairoch, A. 1997. Protein Sequence
Annotation in the Genome Era: The Annotation
Concept of SWISS-PROT + TREMBL. In Proceedings of the
Fifth International Conference on Intelligent Systems for
Molecular Biology, 33–43. Menlo Park, Calif.: AAAI
Press.
Brutlag, D. L. 1998. Genomics and Computational
Molecular Biology. Current Opinion in Microbiology
1(3): 340–345.
Attwood, T. K. 2000. Genomics. The Babel of Bioin-
AI MAGAZINE
Baeza-Yates, R.; and Ribeiro-Neto, B., eds. 1999. Modern Information Retrieval. New York: ACM Press.
Apweiler, R. 2001. Functional Information in SWISSPROT: The Basis for Large-Scale Characterization of
Protein Sequences. Briefings in Bioinformatics 2(1):
9–18.
Ashburner, M.; Ball, C. A.; Blake, J. A.; Botstein, D.;
Butler, H.; Cherry, J. M.; Davis, A. P.; Dolinski, K.;
Dwight, S. S.; Eppig, J. T.; Harris, M. A.; Hill, D. P.; Issel-Tarver, L.; Kasarskis, A.; Lewis, S.; Matese, J. C.;
Richardson, J. E.; Ringwald, M.; Rubin, G. M.; and
Sherlock, G. 2000. Gene Ontology: Tool for the Unification of Biology. The Gene Ontology Consortium.
Nature Genetics 25(1): 25–29.
54
formatics. Science 290(5491): 471–473.
Bork, P.; Ouzounis, C.; Sander, C.; Scharf, M.; Schneider, R.; and Sunnhammer, E. 1992. What’s in a
Genome?” Nature 358(6384): 287.
Carter, P.; Liu, J.; and Rost, B. 2003. PEP: Predictions
for Entire Proteomes. Nucleic Acids Research 31(1):
410–413.
Cowie, J., and Lehnert, W. 1996. Information Extraction. Communications of the ACM 39(1): 80–91.
Dasarathy, B. V. 1991. Nearest-Neighbor (NN)
Norms: NN Pattern Classification Techniques. Washington, D.C.: IEEE Computer Society.
Devos, D., and Valencia, A. 2001. Intrinsic Errors in
Genome Annotation. Trends in Genetics 17(8):
429–431.
Articles
Devos, D., and Valencia, A. 2000. Practical Limits of
Function Prediction. Proteins 41(1): 98–107.
Ding, J.; Berleant, D.; Nettleton, D.; and Wurtele, E.
2002. Mining MEDLINE: Abstracts, Sentences, or
Phrases? Paper presented at the Pacific Symposium
on Biocomputing, 6–10 January, Kapalua, Hawaii.
Doerks, T.; Bairoch, A.; and Bork, P. 1998. Protein Annotation: Detective Work for Function Prediction.
Trends in Genetics 14(6): 248–250.
Eisenberg, D.; Marcotte, E. M.; Xenarios, I.; and
Yeates, T. O. 2000. Protein Function in the Post-Genomic Era. Nature 405(6788): 823–826.
Eisenhaber, F., and Bork, P. 1999. Evaluation of Human-Readable Annotation in Biomolecular Sequence
Databases with Biological Rule Libraries. Bioinformatics 15(7–8): 528–535.
Fleischmann, R. D.; Adams, M. D.; White, O.; Clayton, R. A.; Smith, H. O.; and Venter, J. C. 1995.
Whole-Genome Random Sequencing and Assembly
of Haemophilus Influenzae Rd. Science 269(5223):
496–512.
Fleischmann, W.; Moller, S.; Gateau, A.; and Apweiler, R. 1999. A Novel Method for Automatic Functional Annotation of Proteins. Bioinformatics 15(3):
228–233.
Fraser, C. M.; Gocayne, J. D.; White, O.; Adams, M.
D.; Philips, C. A.; and Venter, J. C. 1995. The Minimal Gene Complement of Mycoplasma Genitalium.
Science 270(5235): 397–403.
Friedman, C.; Kra, P.; Yu, H.; Krauthammer, M.;
Rzhetsky, A. 2001. GENIES: A Natural Language–Processing System for the Extraction of Molecular Pathways from Journal Articles. Bioinformatics 17(Suppl.
1): S74–S82.
Gaasterland, T., and Sensen, C. W. 1996. MAGPIE: Automated Genome Interpretation. Trends in Genetics12(2): 76–78.
Galperin, M. Y., and Koonin, E. V. 2000. Who’s Your
Neighbor? New Computational Approaches for
Functional Genomics. Nature Biotechnology 18(6):
609–613.
Harrison, P. M.; Bamborough, P.; Daggett, V.; Prusiner, S.; and Cohen, F. E. 1997. The Prion Folding Problem. Current Opinion in Structural Biology 7(1): 53–59.
Hatzivassiloglou, V.; Duboue, P. A.; and Rzhetsky, A.
2001. Disambiguating Proteins, Genes, and RNA in
Text: A Machine Learning Approach. Bioinformatics
17(Suppl. 1): S97–S106.
Hersh, W. R.; Evans, D. A.; Monarch, I. A.; Lefferts, R.
G.; Handerson, S. K.; and Gorman, P. N. 1992. Indexing Effectiveness of Linguistic and Nonlinguistic Approaches to Automatic Indexing. Amsterdam, The
Netherlands: Elsevier Science.
Iliopoulos, I.; Enright, D. A.; and Ouzounis, C. A.
2001. TEXTQUEST: Document Clustering of MEDLINE
Abstracts for Concept Discovery in Molecular Biology. Paper presented at the Pacific Symposium on Biocomputing, 3–7 January, Mauna Lani, Hawaii.
Junker, V.; Contrino, S.; Reischmann, W.; and Apweiler, R. 2000. The Role SWISS-PROT and TREMBL Play
in the Genome Research Environment. Journal of
Biotechnology 78(3): 221–234.
Karp, P. D.; Riley, M.; Paley, S. M.; Pellegrini-Toole, A.;
and Krummenacker, M. 1999. Eco CYC: Encyclopedia
of Escherichia coli Genes and Metabolism. Nucleic
Acids Research 27(1): 55–58.
Koonin, E. V. 2000. Bridging the Gap between Sequence and Function. Trends in Genetics 16(1): 16.
Krauthammer, M.; Kra, P.; Iossifov, I.; Gomez, S. M.;
Hripcsak, G.; Hatzivassiloglou, V.; Friedman, C.; and
Rzhetsky, A.. 2002. Of Truth and Pathways: Chasing
Bits of Information through Myriads of Articles.
Bioinformatics 18(Suppl. 1): S249–S257.
Krawiec, S., and Riley, M. 1990. Organization of the
Bacterial Chromosome. Microbiology Review 54(4):
502–539.
Kretschmann, E.; Fleischmann, W.; et al. 2001. Automatic Rule Generation for Protein Annotation with
the C4.5 Data-Mining Algorithm Applied on SWISSPROT. Bioinformatics 17(10): 920–926.
Lewis, D. D., and Ringuette, M. 1994. Comparison of
Two Learning Algorithms for Text Categorization.
Paper presented at the Third Annual Symposium on
Document Analysis and Information Retrieval
(SDAIR’94), 11–13 April, Las Vegas, Nevada.
Lewis, S.; Ashburner, M.; and Reese, M. G. 2000. Annotating Eukaryote Genomes. Current Opinion in
Structural Biology 10(3): 349–354.
Liu, J., and Rost, B. 2001. Comparing Function and
Structure between Entire Proteomes. Protein Science
10(10): 1970–1979.
Lodish, H.; Berk, A.; Baltimore, D.; and Darnell, J.
2000. Molecular Cell Biology. New York: Freeman.
Luscombe, N. M.; Greenbaum, D.; and Gerstein, M.
2001. What Is Bioinformatics? A Proposed Definition
and Overview of the Field. Methods of Information in
Medicine 40(4): 346–358.
Mewes, H. W.; Frishman, D.; Gruber, C.; Stocker, S.;
and Weil, B. 2000. MIPS: A Database for Genomes
and Protein Sequences. Nucleic Acids Research 28(1):
37–40.
Nair, R., and Rost, B. 2002a. Inferring Subcellular Localization through Automated Lexical Analysis.
Bioinformatics 18(Suppl. 1): S78–S86.
Nair, R., and Rost, B. 2002b. Sequence Conserved for
Subcellular Localization. Protein Science 11(12):
2836–2847.
Ng, S. K., and Wong, M. 1999. Toward Routine Automatic Pathway Discovery from Online Scientific Text
Abstracts. Genome Informatics Series Workshop.
Genome Informatics 10(1): 104–112.
Ohta, Y.; Yamamoto, Y.; Okazachi, T.; Uchiyama, I.;
and Takagi, T. 1997. Automatic Construction of
Knowledge Base from Biological Papers. Paper presented at the International Conference of Intelligent
Systems in Molecular Biology, 21–25 June, Halkidiki,
Greece.
Ouzounis, C.; Casari, G.; Sander, C.; Tamames, J.; and
Valencia, A. 1996. Computational Comparisons of
Model Genomes. Trends in Biotechnology 14(B):
280–285.
Ouzounis, C.; Perez-Irratxeta, C.; Sander, C.; and Valencia, A. 1998. Are Binding Residues Conserved? Paper presented at the Pacific Symposium on Biocom-
SPRING 2004
55
Articles
puting, 4–9 January, Kapalua, Maui,
Hawaii.
Overbeek, R.; Larsen, N.; Smith, W.; Maltsev, N.; and Selkov, E. 1997. Representation
of Function: The Next Step. Gene 191(1):
GC1–GC9.
Pruess, M.; Fleischmann, W.; Kanapin, A.;
Servant, F.; and Apweiler, R. 2003. The Proteome Analysis Database: A Tool for the In
Silico Analysis of Whole Proteomes. Nucleic
Acids Research 31(1): 414–417.
Riley, M. 1993. Functions of the Gene Products of Escherichia coli. Microbiology Review
57(4): 862–952.
Riley, M., and Labedan, B. 1997. Protein
Evolution Viewed through Escherichia coli
Protein Sequences: Introducing the Notion
of a Structural Segment of Homology, the
Module. Journal of Molecular Biology 268(5):
857–868.
Rost, B. 2002. Enzyme Function Less Conserved Than Anticipated. Journal of Molecular Biology 318(2): 595–608.
Rost, B. 1999. Twilight Zone of Protein Sequence Alignments. Protein Engineering
12(2): 85–94.
Rost, B., and Sander, C. 1996. Bridging the
Protein Sequence-Structure Gap by Structure Predictions. Annual Review of Biophysics and Biomolecular Structure 25:113–
136.
Salton, G. 1989. Automatic Text Processing.
Reading, Mass.: Addison-Wesley.
Schutze, H.; Hull, D. A.; and Pederson, J. O.
1995. A Comparison of Classifiers and Document Representation for the Routing
Problem. In Proceedings of the Ann Int
ACM SIGIR Conference on Research and
Development in Information Retrieval (SIGIR ‘95), 229–237. New York: Association
of Computing Machinery.
Shah, I., and Hunter, L. 1997. Predicting
Enzyme Function from Sequence: A Systematic Appraisal. In Proceedings of the Fifth
International Conference on Intelligent Systems for Molecular Biology, 276–283. Menlo
Park, Calif.: AAAI Press.
Shatkay, H.; Edwards, S.; William, W. J.; and
Boguski, M. 2000. Genes, Themes, and Microarrays: Using Information Retrieval for
Large-Scale Gene Analysis. Paper presented
at the International Conference on Intelligent Systems in Molecular Biology, 19–23
August, La Jolla, California.
Stapley, B. J., and Benoit, G. 2000. Biobibliometrics: Information Retrieval and Visualization from Cooccurrences of Gene
Names in Medline Abstracts. Paper presented at the Pacific Symposium on Biocomputing, 4–8 January, Honolulu, Hawaii.
Stapley, B. J.; Kelley, L. A.; and Sternberg,
M. J. 2002. Predicting the Subcellular Location of Proteins from Text Using Support
56
AI MAGAZINE
Vector Machines. Paper presented at the Pacific Symposium on Biocomputing, 3–7
January, Lihue, Hawaii.
Stephens, M.; Palakal, M.; Mukhopadhyay,
S.; Raje, R.; and Mostata, J. 2001. Detecting
Gene Relations from MEDLINE Abstracts.
Paper presented at the Pacific Symposium
on Biocomputing, 3–7 January, Honolulu,
Hawaii.
Stevens, R.; Goble, C. A.; and Bechhofer, S.
2000. Ontology-Based Knowledge Representation for Bioinformatics. Briefings in
Bioinformatics 1(4): 398–414.
Tamames, J.; Ouzounis, C.; Casari, G.;
Sander, C.; and Valencia, A. 1998. EUCLID:
Automatic Classification of Proteins in
Functional Classes by Their Database Annotations. Bioinformatics 14(6): 542–543.
Thomas, J.; Milward, D.; Ouzounis, C.; Pulman, S.; and Carroll, M. 2000. Automatic
Extraction of Protein Interactions from Scientific Abstracts. Paper presented at the Pacific Symposium on Biocomputing, 4–8
January, Honolulu, Hawaii.
Todd, A. E.; Orengo, C. A.; and Thornton, J.
M. 2001. Evolution of Function in Protein
Superfamilies, from a Structural Perspective. Journal of Molecular Biology 307(4):
1113–1143.
Tsoka, S., and Ouzounis, C. A. 2000. Recent
Developments and Future Directions in
Computational Genomics. Federation of European Biochemical Societies Letters 480(1):
42–48.
Valencia, A. 2002. Search and Retrieve:
Large-Scale Data Generation Is Becoming
Increasingly Important in Biological Research. But How Good Are the Tools to
Make Sense of the Data? European Molecular
Biology Organization (EMBO) Reports 3(5):
396–400.
Valencia, A., and Pazos, F. 2002. Computational Methods for the Prediction of Protein Interactions. Current Opinion in Structural Biology 12(3): 368–373.
Webb, E. C. 1992. Enzyme Nomenclature
1992. Recommendations of the Nomenclature
Committee of the International Union of Biochemistry and Molecular Biology. San Diego,
Calif.: Academic.
Wilson, C. A.; Kreychman, J.; and Gerstein,
M. 2000. Assessing Annotation Transfer for
Genomics: Quantifying the Relations between Protein Sequence, Structure, and
Function through Traditional and Probabilistic Scores. Journal of Molecular Biology
297(1): 233–249.
Wrzeszczynski, K. O., and Rost, B. 2003.
Cataloguing Proteins in Cell Cycle Control.
In Cell Cycle Checkpoint Control Protocols, ed.
H. Lieberman, 219–233. Totowa, N.J.: Humana.
Yakushiji, A.; Tateisi, Y.; Miyao, Y.; and Tsu-
jii, J. 2001. Event Extraction from Biomedical Papers Using a Full Parser. Paper presented at the Pacific Symposium on Biocomputing, 3–7 January, Honolulu,
Hawaii.
Yang, Y. 1996. An Evaluation of Statistical
Approaches to MEDLINE Indexing. Paper
presented at the Conference of the American Medical Informatics Association, 26–29
October, Washington, D.C.
Yang, Y., and Chute, C. G. 1992. An Application of Least Squares Fit Mapping to Clinical Classification. Paper presented at the
Annual Symposium on Computer Applications in Medical Care, 2–6 November,
Washington, D.C.
Yang, Y., and Liu, X. 1999. A Reexamination of Text Categorization Methods. In
Proceedings of the ACM SIGIR Conference
on Research and Development in Information Retrieval (SIGIR’99), 42–49. New York:
Association of Computing Machinery.
Yang, Y., and Pederson, J. P. 1997. A Comparative Study on Feature Selection in Text
Categorization. Paper presented at the
Fourteenth International Conference on
Machine Learning, 8–12 July, Nashville,
Tennessee.
Rajesh Nair is a graduate
student in the Physics
Department at Columbia
University. He has been
responsible for developing a number of widely
used
bioinformatics
tools for the prediction
of protein subcellular localization such as PREDICTNLS, LOCKEY, LOCHOM, and LOC3D. His recent research interests have focused on the application of
machine learning techniques such as neural networks and support vector machines
to predicting various aspects of protein
function. His e-mail address is nair@cubic.bioc.columbia.edu.
Burkhard Rost is an associate professor of biochemistry and molecular
biophysics at Columbia
University. He was responsible for creating
the PREDICT PROTEIN server
for protein secondary
structure prediction. PREDICT PROTEIN is the most widely used public
server for protein structure prediction and
has handled over a million requests. His research focuses on the prediction of protein
structure and function by combining
means from simple statistics to AI and evolutionary information. His e-mail address is
rost@columbia.edu.